advanced process control and global optimization for ... outline control and optimization of complex...
TRANSCRIPT
Advanced Process Control and Global Optimization for
Complex Processes
Yu Yang
June 2nd, 2016 UTC Institute for Advanced Systems Engineering
University of Connecticut
2
Research Scope
Regulatory Control (PID)
Planning/Scheduling
StrategicPlanning
Customer
Advanced Process Control (APC)
$
$
$
$
Decision Hierarchy in Process Industry
complexity
[year]
[month]
[week]
[min]
[sec]
3
Outline
Control and Optimization of Complex Chemical Processes
Advanced Process Control for PDE System
Global Optimization under Uncertainty
Model Predictive Control for Nonlinear ODE System
4
5
Model Predictive Control
Artificial Pancreas
Semiconductor Manufacturing Pharmaceutical Manufacturing
Refinery
6
Nonlinear Model Predictive ControlOptimization
Objective Function
Model Constraints
min t = k+1
k+p
performance
dx/dt =f(x,u) ymin≤ y≤ymax
FuturePast
Predicted performancew/ optimized decision
Real outcome
k k+1 k+ptime
Decision
Real PlantObserver
Decision
Feedforward
NewInformation
Setpoint
Limitations: 1. No robust stability2. Significant online computational demands
• Control
- Deviation from setpoint
T
setpoint setpoint
k p
t k
x t x Q x t x
7
Robust Stability
Robust control Lyapunov function (RCLF: )
, ,
, , , : bounded uncertainties
x t f x f x t g x g x t u t
f x t g x t
Robustly Invariant Set:
TV x x Px
One-step worst-case MPC:
k k+1 k+ptime
FuturePast
Worst-case model
Nominal model
1V x k V x k
RCLF forms the terminal constraint
,
inf sup 0T
u U f g
x P f x f g x g u
8
Robust Control Lyapunov Function
Objective: Enlarging ROA and reducing residual set
ˆ
,
min max
ˆ
min max
ˆ ˆmax
s.t. inf sup 0
0
ˆ
ˆ ˆ max
T
x
TP
x
T
u f g
T T
x
x Px
x Px
x P f x f g x g u
u u u
P
x
x Px x Px
Bilevel Optimization:Fractional programming: Dinkelbach’s methodCoordinate search: Single variable optimization
Yang & Lee, IET control theory and application, 2012Yang & Lee, Journal of Process Control, 2011
Characterize the level of ROA
Characterize the level of residual set
9
-
-
-
-
-
Performance Improvement
Initial Control
Policies
Value Function
Approximation
1( ) min ( , ) ( ( , ))i i
uJ x E r x u J f x u
Off-line policy improvement
On-line Implementation
PID/One-step worst-case MPC
Approximate Dynamic Programming
H. Nosair, Y. Yang & J. M. Lee, Control Engineering Practice, 2010Y. Yang & J. M. Lee, Computers and Chemical Engineering, 2010
arg min ,u
u k J f x u
T
setpoint setpoint( )k p
t k
J x x t x Q x t x
MPC: , 1 , ,u k u k u k p
10
Example
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2386
388
390
392
394
396
398
400
402
CA
TR
ADP
MPC
ROA
Start
MPC ADP
Averagecomputational time
8.6s 0.1s
Average control cost 252.7 230.2
Y. Yang & J. M. Lee, Journal of Process Control, 2012Y. Yang & J. M. Lee, Journal of Process Control, 2013
/
0 0
/
0 0
1 R
R
E RT
A A A A
E RT
R R R A
p p
FC C C k e C
V
QF HT T T k e C
V C C V
Uncertainty
Uncertainty
T
setpoint setpoint
k p
t k
x t x Q x t x
11
12
PDE Control: Motivation
Heat equation Fluid mechanics Rod pump
Crystallization Lithium battery Engine
13
PDE Control: Application
2 2
2
2 2
, , t ,x z t x z x z tv c
zz t
Sensor: SurfaceEstimation: Downhole load & Displacement
: displacement
: position
: time
x
z
t
Stress difference, Hooke’s law
Acceleration Friction
14
0 1 2 3 4 5 6 7 8 9 100
2
4
6
8
10
12
Time (s)
Dis
pla
cem
ent
(Feet)
Proposed method
Classical mathod
PDE Control: Application
Displacement and load estimation:
0 1 2 3 4 5 6 7 8 9 100.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2x 10
4
Time (s)
Dynam
ic load (
Lbs)
Proposed method
Classical method
0 1 2 3 4 5 6 7 8 9 100
2
4
6
8
10
12
Time (s)
Dis
pla
cem
ent
(Feet)
Classical method
Proposed method
0 1 2 3 4 5 6 7 8 9 100.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2x 10
4
Time (s)
Dynam
ic load (
Lbs)
Classical method
Proposed method
Do
wn
ho
le D
isp
lace
men
t
Do
wn
ho
le L
oad
( Y. Yang & S. Dubljevic, SPE Annual Conference, 2012 )
15
PDE Control: Model Reduction
Spectral method:
My theoretical contributions: Handle the truncation residual in the model predictive control
Controller and observer synthesis by taking residual into account
1
, i i
i
x z t a t z
1
,sn
i i
i
x z t a t z
Truncation
1
, ,sn
i i
i
x z t a t z R z t
State#
FD 2nd order(max error)
FD 4th order(max error)
Fourier spectral(max error)
16 6.13e-1 2.36e-1 2.55e-4
32 1.99e-1 2.67e-2 1.05e-11
64 5.42e-2 1.85e-3 6.22e-13
Spectral Methods in Fluid Dynamics, Canuto, 1988 (Table 1.1 1st order wave equation).
Slow modal state
16
PDE Control: Model Predictive Control
Two-phase flow (linearized K-S equation)
0
2
4
6
8
0
5
10
15
20
-1.5
-1
-0.5
0
0.5
1
1.5
Time (s)
x(
,t)
4 2
4 2
1 2
0,
0, 0, , ,
0, 0, , ,
, : boundary inputs,
, : variance of the thin film thickness.
x x xv
t
x t x l t d u t d t
x t x l t u t
u t t
x t
(Y. Yang & S. Dubljevic, Journal of process control, 2013)
Spectral methodGalerkin’s model reductionMPC ( Full state feedback )
0 5 10 15 20-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
Time (s)
y
0.5 0.6 0.7 0.8 0.9 1-0.56
-0.55
-0.54
-0.53
-0.52
With fast modal output bound
Without fast modal output bound
17
PDE Control: Controller & Observer
Controller and observer synthesis:
1 2 1 2 1 2, , , , , , , , , , u u u u s u s u sn n n n n n n n na a a a a a a a a
1. How to find the order of slow modal states?2. How to design the feedback law and observer gains to obtain a closed-loop
system with large enough region of attraction?
Questions
Slow modal states
ˆ ˆ ˆ1
1
ˆ1
u u u u s u
s s s s
u
u uu
a k a k u k y k y k y k
a k a k u k
u k
L
K Ga k u k
B
B
18
PDE Control: Controller & Observer
Results (2nd order Parabolic PDE Stabilization)
00.2
0.40.6
0.81
02
46
810
-100
-50
0
50
100
zt
x
(Y. Yang & S. Dubljevic, European journal of control, 2014)
2
2
1
0
, , , ,
0,
1, 0
, 1
0,1
x x xz t b z t c z t dx z t
t zz
x t u t
xt
z
y t x z t z dz
z
ObserverSpectral methodGalerkin’s model reductionOutput feedback controller (LMI)
Boundary sensor
19
20
Motivation
Metabolic network RefineryWater network
Global Optimization: mixed-integer nonlinear programming (MINLP)
T
,min
s.t. ,
, ,
, 0,1 .
x yc x
Ax b
F x y d
x y
21
Objective: decide the optimal crude oil procurement and refinery operations to maximize the profit, meet market demands and satisfy quality specifications.
Overview
Refinery Optimization under Uncertainty
Market Demand
Profit
Quality
(U.S. Patent Pub. No: US2016/0140448 A1)
22
Flowchart
Model 1
Blending & Splitting (Pooling problem)
Vapor pressure
Octane number
Sensitivity
Sulfur
Viscosity
Multi-mode
J. P. Favennec, Refinery operation and management, Editions TECHNIP, 2001
z x y
23
Uncertainties:
Model 1: Uncertainties
Crude oil yields: www.bp.com
Crude VR yield Sulfur
Skarv 10.3% 0.209%
Saturno 26.7% 0.373%
Plutonio 21.6% 0.2%
Brent 13.6% 0.253%
Hungo 25.5% 0.321%
Polvo 37.2% 0.741%
Zakum 10.9% 0.966%
Basra 26.2% 1.826%
Thunder 20.1% 0.416%
Mars 27.6% 1.216%
Gaussian Distribution: Mean=nominal valueStandard deviation=0.1×nominal value
0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.240
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Crude1 Sulfur wt%
Pro
babili
ty
=0.0157
Sampling: 120 scenarios with
different VR yield & sulfur content
Scenario 1 VR Sulfur
… …Scenario 2 VR Sulfur
… …Scenario 3 VR Sulfur
… …Scenario 4 VR Sulfur
… …Scenario 5 VR Sulfur
… …
24
Solution: Stochastic Programming
Stochastic programming vs. Deterministic method:
Crude oil
Selection
&
purchase
Crude oil
real quality
(TBP)
Refinery operations:
Cut point temperature,
Flow rate
UncertaintiesStage I Stage II
Deterministic method:
*
Nominalmax Cost Profitd
d d dx
x x x
Stochastic method:
* max Cost Profit Prs
s s i s ix
i
x x x
Pr : Probability of scenario
: Decison of deterministic method
: Decison of stochastic method
i
d
s
i
x
x
1Pr
2Pr
3Pr
4Pr
25
min cost of crude oil purchase product sales operation costs
s.t. Crude oil source restrictions, Material balance,
Unit capacity limits,
Demand requirements,
Quality constraints,
Pooling equations
Two-stage Stochastic Formulation (Model 1):
Solution: Stochastic Programming
10 Crude oil
120 Scenarios: different VR yields and sulfur concentration
Discretization of each crude oil quantity:
5000 barrels = 1 lot
We use the non-convex generalized Benders decomposition (NGBD).
Model (120 scenarios):
13061 constraints,
100 binary variables
13800 continuous variables
2760 bilinear terms
( Yang & Barton, AIChE Journal, 2016 )
26
Method: NGBD
Algorithm:
Cutting plane Cutting plane Cutting plane
120
1
min cost of crude oil purchase+ Scenario cost /120
s.t. Crude oil source restrictions
i
i
Crude Crude
1min Scenario
s.t. Scenario constraints
Crude
2min Scenario
s.t. Scenario constraints
120min Scenario
s.t. Scenario constraints
Lower bound solution: MILP
Upper bound solution: NLP Relaxation: LP
27
Results: Profit/Time
Profit/month (18% improvement)
Computational Time
Million $
13.514
14.515
15.516
16.517
17.518
18.5
Stochastic Deterministic
Average profit
GAMS1 C++2 (BARON, Antigone)
Total time 12.2 hours 10 mins No solution in 1 day
1. General Algebraic Modeling System (GAMS)2. Joint work with Rohit Kannan.
28
Model 2: Flowchart
Refinery flowchart
J. P. Favennec, Refinery operation and management, Editions TECHNIP, 2001
29
Decision variables: cut point temperature
Model 2: CDU Model
dis,VR
bot,VR
: Total mole flowrate in the distillate of VR cut
: Total mole flowrate in the bottom (yield) of VR cut.
P
P
,j j jT T T
Crude oil TBP Pseudocomponent
bot, j dis, j
FI equation
, P P
Vapor pressure
Equilibrium ratio
A. M. Alattas, I. E. Grossmann and I. Palou-Rivera, I&EC research, 2011, 2012.
30
Model 2: CDU Model
Fractionation index (FI) equation
dis, ,
,
bot , ,
dis, ,
bot, ,
, ,ini
,
: Mole fraction of pseudocomponent in the distillate of cut
: Mole fraction of pseudocomponent in the bottom of cut
if
FIj i
j i j
j i
j i
j i
r j j i j
s j
xK T
x
x i j
x i j
FI T Tb TFI
FI
,end
, : true boiling point of pseudocomponent if i
j i j
Tb iT Tb T
, s,dis, ,
, ,
bot , ,
,
1 , 0,1 ,
VP, VP : vapor pressure, : equilibrium constant
r j jFI FIj i
j i j i j i j i i
j i
i j
j i j
xK T v K T v v
x
TK T K
P
Gilbert, R.J.H., AIChE Journal, 1966.
31
Model 2: CDU Model
Vapor pressure surrogate model
0.65 0.7 0.75 0.8 0.85 0.9 0.950
0.5
1
1.5
2
2.5
3
3.5
4
Tr
Vapor
Pre
ssure
(atm
)
State equation data
Approximation data
2
, , ,
, 2 , 1 , 0
j i j i j i
j i j i j iVP Tr Tr Tr
,
1.5 3 6
, , , ,
,
1.5 3 6
, , , ,
,
,
VP
5.96 1 1.18 1 0.56 1 1.32 1exp
4.79 1 0.41 1 8.91 1 4.98 1exp
where , : cut,
: pseudocompone
j i
j i j i j i j i
j i
j i j i j i j i
j i
j
j i
ci
Tr
Tr Tr Tr TrPc
Tr
Tr Tr Tr Tr
Tr
TTr j
T
i
nt, : parameterPc
,j j jT T T
32
0 20 40 60 80 100 120300
400
500
600
700
800
900
1000
1100
Vol%
Tem
pra
ture
(K
)
Scenario1 Crude 1
Scenario2 Crude 1
Uncertainty is in the heavy part of true boiling point (TBP) curve
Model 2: Uncertainty
4
4 types of crude oil,
2 TBP curves for each:
2 scenarios
33
Two-stage Stochastic Formulation (Model 2):
Solution: Stochastic Programming
4 Crude oil
Discretization of each crude oil
quantity:
5000 barrels = 1 lot
Model (16 scenarios):
10058 constraints,
168 binary variables (128 in CDU model)
8452 continuous variables
8032 bilinear terms
3456 signomial terms:
min cost of crude oil purchase operation costs product sales
s.t. Crude oil source restrictions, Material balance,
Unit capacity limits
,
CDU model,
Demand requirements,
Quality constraints,
Pooling equations
7
jK T 6.9
jK T 4.25
jK T
34
Nominal Scenario
Difficult global optimization problem (ANTIGONE)
0 500 1000 1500 2000 2500 3000 3500 4000 45000
1
2
3
4
5
6
7
8
9
10
Time (s)
Rel
ativ
e ga
p (%
)
Relative gap
Upper bound-Lower bound=Relative gap=
Upper bound
(ANTIGONE: global optimization software, Misner & Floudas)( Yang & Barton, ADCHEM, 2015 )
Preprocessing:Interval reduction
35
Optimization: NGBD
Initialization
LBD>=UBDPB?
Primal bounding problem (LP)
PBP feasible?
Feasibility problem (LP)
Relaxed master problem (MILP)
Dual-PBP-PCR (MILP)
Update: piece, UBDPB
Update LBD
Primal problem (MINLP)
Update: UBD, UBDPB
UBDPB>=UBD
End
New integer realization
Feasibility cut
Optimality cuts: PBP cut, Dual-PBP-PCR cut
Yes
No
Yes
No
Yes
No
\U
UBDPB= min obj Dual-PBP-PCRk
k T
New Features:1. Enhanced interval reduction2. Adaptive piecewise convex relaxation3. Local dual cutting plane4. Capable to handle non-separable case
36
Comparison
Computational time
Time/Iterations Adaptive1 Non-adaptive2
PBP (s) 3932 3896
Dual-PBP-PCR (s) 41572 87009
PP (s) 10686 6094
RMP (s) 148 73
Total time (h) 15.6 26.9
Iteration 126 95
1. Add binary variable if relaxation gap >=10%.2. Use 6 binary variables for piecewise relaxation .
37
Results:
Profit distribution ($/barrel)
4.5 4.6 4.7 4.8 4.9 5 5.1 5.2 5.3 5.4 5.50.12
0.14
0.16
0.18
0.2
0.22
0.24
0.26
Profit $/barrel
Pro
babi
lity
Deterministic
Stochastic
Deterministic
expected profit
Stochastic
expected profit
38
Summary
Research Scope: Advanced process control/Scheduling for complex processes under uncertainties
Methods: MPC, reinforcement learning, spectral method, two-stage stochastic programming (NGBD).
Applications: Rod pump, refining process design,
CSTR control
39
Supplement
Yu Yang
Process Systems Engineering LaboratoryDepartment of Chemical Engineering
Massachusetts Institute of Technology
40
PDE Control: Model Reduction
Parabolic PDE (Example: 2nd order)
Eigenfunction analysis 2
2: b c d
zz
A
1
Solve
Let , then
,s
i i
i i
n
i i
i
z z
f z z
x z t a t z
A
2
2
1
0
, , ,
, c c
x z t x z t x z tb c dx
t zz
y t x z t z z dz
0
1,0, , 0
,0
15
x tx t u t
z
x z x t
u t
0 20 40 60 80 100 120 140 160 180 200-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
Space step
i(z
)
Initial condition
Boundary condition
Input constraintPoint-wise output
41
1
1
1 1
uu
u u
u su s
u
nn
n n
s
n nn n
f
Bdza
Bdz
Bdza u
a
Bdz
aBdz
PDE Control: Model Reduction
Decomposition & Boundary transformation:, , ,
: unstable modal state
: stable slow modal state
: stable fast modal state
u s f
u
s
f
a a a a
a
a
a
1, , 0,un 1, , 0
un is boundedfa t
42
PDE Control: MMPC
Modal model predictive control:
(Assume can be measured)
T T T
1 2,
0
min
ft
s su
t
y Qy u W u W
min max ,s f uy y y y y
1 2f f f f fy Ca B u B
1 2s.t. ,s s s s sa a B u B
7, , ,s uu u a a
1 2 ,u u u u ua a B u B
1 2 ,s s s s sy Ca B u B
1 2 ,u u u u uy Ca B u B
,x z t
• Static bound of the output for fast modal states
• Output for unstable modal states
• Control performance
,cy x z t output
,u sa a are known in real time
43
PDE Control: Controller & Observer
Dynamic upper and lower bounds for: fy k
1 1 1f f fy k y k y k
k
y fy k
fy k
fy k
44
PDE Control: Controller & Observer
Solve linear matrix inequality (LMI):
T
1
T
2
2
max
T T
1 1 2 2
0 ,
0 0,
,
0, 0.
u u
u
u u u
S K G
S K
K G K u I
S S S S
Convergence:
LMI is guaranteed to be feasible by increasing
the dimension of sa
lim , 0.k
x z k
Feasibility:
, ,u u uK G L
1
0, 0
0 1
0 ,
,
u u u
u u u
u u u u
u u u
K G K
L L
K G K
B B
C
1
0, 0
0 1
0 ,
,
u u u
u u u
u u u u
u u u
K G K
L L
K G K
B B
C
where
45
Model 2: CDU Model
Pseudocomponent method:
Distillate
BottomjT
jT
Bo
ilin
g Te
mp
era
ture
Pseudocomponent
Light key component
Heavy key component
46
Optimization: Range Reduction
Upper bound solution
continuousx
ObjectiveFunction
47
Optimization: Relaxation
Convex/Piecewise convex relaxation
PPobj
c
PBPobj
c
Convex relaxation gap
Piecewise convex relaxation gap
Partition point
48
Optimization: Piecewise Relaxation
How to select the partition variable/point to get tighter cutting plane?
Partition point
Relaxation
If violation ratio
lo lo lo lo
up up up up
lo up lo up
up lo up lo
u xP u x P xP x P
u x P xP x P
u x P xP x P
u x P xP x P
u
threshold,
then one variable in this term is partitioned.
xP
xP
49
Optimization: Local Cutting Plane
Objective: design the dual cutting plane only valid in a specific region .
Benefits:
1. Better bounds for convex relaxation.
2. Extra tight constraints for Dual-PBP-PCR.
,p pc c
, , , , ,, , , p i p i p p p p p i p i p p i pf y c c c c f y c y c
, ,, , , p i p p p p p i p p
i i
f c c c c f c c
91
,
1
Note: 20 2t
p p p t p
t
c d D Q
50
9* 1 *
, , , ,, , , ,
1
, , , , , ,
, , , , , , , , , , , , , , ,
, , ,
min 20 2
s.t. , , , 0, , ,
,
t
s s p p k s p p t px y f b
p k t
s p j s p i s s p k s p p
k
p i j s p j s p i s p j s p i s p j s p i s
p i j
w x f d D Q
G x y f f c c
b f y f y f y
b
, , , , , , , , , , , ,
, , , , , , , , , , , , , , ,
, , , , , , , , , , , , , , ,
, , , , , , , , , ,
,
,
,
,
s p j s p i s p j s p i s p j s p i s
p i j s p j s p i s p j s p i s p j s p i s
p i j s p j s p i s p j s p i s p j s p i s
p i j s p j s p i j s p i
i
f y f y f y
b f y f y f y
b f y f y f y
b f b f
, , , ,
,
, , , 0,1 ,
, , .
s
j
n
s p i s p j s s sx y f
i j p
Optimization: Local Cutting Plane
Enhanced PBP-PCR/Dual-PBP-PCR
, , , ,
9* 1 *
, , ,
1
, ,
, , , ,
, , , , , , , , , , , , , , ,
, , ,
min
s.t. 20 2 ,
, ,
, , , 0,
,
sx y f b
t
p k s p p t p
k t
p k s p p
k
s p j s p i s s
p i j s p j s p i s p j s p i s p j s p i s
p i j
w x
f d D Q
f c c
G x y f
b f y f y f y
b
, , , , , , , , , , , ,
, , , , , , , , , , , , , , ,
, , , , , , , , , , , , , , ,
, , , , , , , , , ,
,
,
,
,
s p j s p i s p j s p i s p j s p i s
p i j s p j s p i s p j s p i s p j s p i s
p i j s p j s p i s p j s p i s p j s p i s
p i j s p j s p i j s p i
i
f y f y f y
b f y f y f y
b f y f y f y
b f b f
, , , ,
,
, , , 0,1 ,
, , .
s
j
n
s p i s p j s s sx y f
i j p
51
Old cut (globally valid):
New cut (local valid):
Optimization: Local Cutting Plane
4 4 9
* T * 1
Dual-PBP-PCR , , ,
1 1 1
obj 20 2t
s s p s p p p s p p t
p p t
c Bc BQ d BQ D
* *
, ,
4 4 9* T * 1
E-Dual-PBP-PCR , , ,
1 1 1
4 4
, ,
1 1, 0 , 1
obj 20 2
1
p h p h
t
s s p s p p p s p p t
p p t
s p h s p h
p ih D h D
c Bc BQ d BQ D
M D M D
0, can be determined systematicallysM * *
E-Dual-PBP-PCR Dual-PBP-PCRNote: obj objc c
Variables: ,d D
52
Optimization: Non-separable case
Bilinear term=stage I variable ×stage II variable
130132
134136
138
106108
110112
114
-3.88
-3.86
-3.84
-3.82
-3.8
-3.78
-3.76
-3.74
x 105
Crude4Crude
1
obj L
BP
LBP1LBP2 & LBP2 Cut
LBP1 cut
LBP1: convex relaxationLBP2: Reformation + convex relaxation